vllm.v1.kv_offload.tiering.p2p.data.nixl
¶
NixlTransport: Data-plane transport for RDMA-based KV block transfers via NIXL.
Threading: the agent is reached from the scheduler thread and, when the tiering manager runs one, from its control-plane thread. The manager's lock serializes them strictly -- one Python thread inside any agent call at a time, with a full memory barrier between.
That lock is the only protection, not a belt over NIXL's own. nixl_agent picks
NIXL_THREAD_SYNC_STRICT only when enable_listen is set, which this transport
does not set, so the agent runs with NIXL_THREAD_SYNC_NONE -- under which NIXL's
internal agent lock compiles down to no-ops, leaving every agent entry point
unguarded. So the manager lock must cover all of them, and the lifetime of every
handle passed between calls. The exception is peer registration moved onto a
worker thread, as NixlConnector's handshake executor does: it runs outside the
lock and relies, like that connector, on NIXL tolerating it beside transfers.
Serialized use from a second thread is supported (checked against NIXL 1.4.1
sources): NIXL never creates its UCX worker in UCS_THREAD_MODE_SINGLE, the only
creator-thread-affine mode. The weakest mode it ever requests is
UCS_THREAD_MODE_SERIALIZED, documented as "multiple threads can access, but only
one at a time", and because enable_prog_thread defaults on here it actually
requests UCS_THREAD_MODE_MULTI and fails construction if the UCX build cannot
provide it.
Two caveats. NIXL binds a shared UCX worker per calling thread
(tlsSharedWorkerMap), which is harmless only while there is exactly one
shared worker -- true here, but it would stop being true if a num_workers
UCX backend param were ever set. And NIXL 1.4.1 does accept an explicit
sync_mode (including NIXL_THREAD_SYNC_RW), which would make its own locking
real; that parameter does not exist before 1.x, so passing it would need a
version gate, the way capture_telemetry does.
Classes:
-
NixlTransport–Manages a NIXL agent, memory registration, and block transfers.
NixlTransport
¶
Bases: DataTransport
Manages a NIXL agent, memory registration, and block transfers.
Wraps the NIXL C library behind a Python interface so the rest of the P2P tier code never touches NIXL types directly. Tracks inflight handles internally and returns completed/failed tags on poll.
Methods:
-
cancel–Cancel inflight transfers by their IDs.
-
poll–Poll inflight transfers.
-
write_blocks–Submit a WRITE transfer to peer_id.
Source code in vllm/v1/kv_offload/tiering/p2p/data/nixl.py
72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 | |
cancel(transfer_ids, mode='immediate')
¶
Cancel inflight transfers by their IDs.
See DataTransport.cancel for the contract. In "wait" mode,
transfers whose release_xfer_handle raises (NIXL could not
complete the abort because the backend is still draining) stay
in self._inflight so a later poll() will observe them.
Source code in vllm/v1/kv_offload/tiering/p2p/data/nixl.py
poll(peer_id=None)
¶
Poll inflight transfers.
When peer_id is given, only transfers submitted for that peer_id are checked and drained — the transport is shared across peer sessions, so an unscoped poll by one session would consume and discard siblings' completions. peer_id None polls every peer (shutdown drain only).
Returns PollResult(done=..., failed=...) with transfer IDs. Completed handles are released automatically.
Source code in vllm/v1/kv_offload/tiering/p2p/data/nixl.py
write_blocks(peer_id, local_idxs, remote_idxs)
¶
Submit a WRITE transfer to peer_id.
Returns a transfer ID, or None if the peer is not registered. The ID is returned via poll() when the transfer completes or fails.
Source code in vllm/v1/kv_offload/tiering/p2p/data/nixl.py
_Inflight
¶
Bases: NamedTuple
A submitted-but-not-yet-drained transfer.
peer_id lets poll() scope to a single owning session, since the
transport is shared across all peer sessions of the engine.